This repository has been archived by the owner on Aug 5, 2022. It is now read-only.

XavierFiller and MSRAFiller fail to fill a bias blob (Duplicate of upstream issue that is fixed) #219

Closed

TechnikEmpire opened this issue Jun 12, 2018 · 4 comments

TechnikEmpire commented Jun 12, 2018

See BVLC/caffe#6275 and BVLC/caffe#6278, your fork suffers from this. As such, things like the well known Yahoo NSFW model will not load.

TechnikEmpire changed the title ~~XavierFilller and MSRAFiller fail to fill a bias blob (Duplicate of upstream issue that is fixed)~~ XavierFiller and MSRAFiller fail to fill a bias blob (Duplicate of upstream issue that is fixed)

TechnikEmpire mentioned this issue

if or not we need to delete the weight_filter{} and bias_filter in the deploy.prototxt? yahoo/open_nsfw#71

Open

Contributor

ftian1 commented Jun 12, 2018

thanks for inform us about this fix. we will review and integrate it if there is no problem.

Author

TechnikEmpire commented Jun 13, 2018

Actually, while this fix enables me to avoid the exception, I'm encountering something else affecting all trained models now.
[libprotobuf FATAL \google\protobuf\repeated_field.h:1431] CHECK failed: (index) < (current_size_)

Regardless of the selected engine (CAFFE||MKLDNN||etc), when doing classification, the output layer values can arbitrarily change per-session. I'm admittedly over my head at this point, but I have a hunch this is related.

Contributor

ftian1 commented Jul 17, 2018

the check failure you saw is not about this fix. it means your caffemodel is not compatible with intel caffe.

ps: the xavier filler fix would be scheduled to future release.

Contributor

ftian1 commented Nov 28, 2018

close it first.

ftian1 closed this as completed

Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.